Privacy-Preserving MRI Data Harmonization for Black-box Models

ICLR 2026 Conference Submission18639 Authors

Keywords: Harmonization, MRI, Black-box Model, Privacy-preserving, Domain Shift Reduction, Disentangled representation, Bayesian optimization

TL;DR: We propose BboxHarmony, a privacy-preserving MRI harmonization framework that adapts black-box models without requiring data sharing or access to model parameters

Abstract: In MRI, variations in scan parameters, sequence, or hardware can lead to discrepancies in image appearance, even for the same subject. These inconsistencies, known as domain shifts, can hinder image analysis and degrade the performance of deep learning models trained on data from specific source domains. MRI harmonization aims to address these issues by aligning target domain images to the source images while preserving anatomical structures. However, most existing harmonization methods require access to both source and target domain data, making data sharing essential and potentially compromising the data privacy that is critical in medical domain. To address this, we propose BboxHarmony, the first harmonization framework tailored for black-box settings, where requires neither data sharing nor access to downstream task model parameters. Our approach estimates the source domain style by searching the manifold of MRI domain style constructed via a disentanglement-based generator using Bayesian optimization guided by black-box model performance. We evaluated our method on brain tissue segmentation task across multiple institutes and demonstrated that it effectively harmonizes target images into source images, leading to improved downstream task performance of a black-box model. By enabling harmonization under strict data-sharing and model-access constraints, BboxHarmony opens an uncharted area of privacy-preserving harmonization in clinical applications.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 18639